Interview
Dan Kokotov: Speech Recognition with AI and Humans | Lex Fridman Podcast #151
- Dan Kokotov is the VP of Engineering at Rev.ai, a company providing human and AI-powered speech-to-text services.
- Rev distinguishes between Rev.com, which handles human transcription and end-user captioning, and Rev.ai, which offers the Automated Speech Recognition (ASR) API and internal ASR tools.
- Rev was founded to improve upon the Upwork model by standardizing verticals (translation and transcription) to remove friction for both clients and freelancers ("Revvers").
- The pricing model for human transcription is $1.25 per minute (originally $1.00), with rush processing options available.
- Automated captioning via Rev.ai costs $0.25 per minute (previously $0.50 or $1.00 depending on the tier).
- Human accuracy in transcription is estimated at 2–3% Word Error Rate (WER).
- Rev.ai currently achieves approximately 14% WER on unstructured audio test suites, competing directly with Google, Amazon, and Microsoft.
- The company operates a "flywheel" business model where human transcriptionists correct ASR drafts, generating high-quality labeled data to train better machine learning models.
- A key strategic focus is leveraging data from the correction process itself (e.g., tracking click sequences and edit timing) to further refine AI, a concept likened to recurrent neural networks learning from human interaction.
- Revvers (freelancers) are primarily based in the US (majority) and other English-speaking countries, given the requirement for strong English proficiency.
- The freelancer demographic is diverse, including stay-at-home parents, students, individuals with social anxiety, and professionals on career breaks.
- Transcribers utilize a specialized editor that starts with an ASR draft; the job involves correcting this text, which is estimated to take 2–3 times the length of the audio in ideal conditions.
- Real-time transcription is currently deemed too difficult for human operators due to the speed of speech, accents, and background noise.
- Rev aims to make audio as searchable and indexable as text, anticipating a future where all meetings and podcasts are automatically transcribed and searchable.
- The company supports programmatic integration via API and third-party tools like Zapier, allowing automatic file transfers to services like Dropbox.
- Lex Friedman (host) and Dan Kokotov discuss the importance of a "creator's love" for a product, citing Steve Jobs and Elon Musk as leaders who obsess over details and resist purely metric-driven decision-making.
- Both guests express concern over platform censorship and the "outrage economy," arguing that optimizing for long-term user happiness and intellectual health is superior to short-term engagement metrics.
- Friedman notes the potential of podcast transcripts to solve the "search and discovery" problem currently limiting the medium.
- Buzzsprout is mentioned as a platform attempting to standardize RSS feed extensions for transcript inclusion.
- The conversation touches on the philosophical implications of the Dune series, specifically the idea that suffering and pressure are necessary to prevent civilizational stagnation.
- Kokotov recommends reading the original six-book Dune sequence by Frank Herbert, while advising against the later books written by his son.
- Recommended works include Aldous Huxley's Brave New World, George Orwell's 1984, and the film Brazil for its depiction of bureaucratic incompetence.
- Kokotov reflects on the transition from individual contributor (programmer) to management, noting the loss of immediate dopamine from coding versus the "ripple effect" of team success.
- Management philosophy at Rev follows "managed by exception," tailoring feedback and motivation to the specific psychological profile of each employee.
- Kokotov views the meaning of life as the drive to create and contribute to humanity, drawing parallels between writing code and biological reproduction.